Decide the AgentController skill taxonomy and promotion pipeline

[justsml]

Question

How should generalized research methodology be extracted from campaign evidence, reviewed, tested, versioned, promoted, discovered, and retired as AgentController-visible Mastra Workspace skills under sandbox/skills? Decide when to deepen existing skills versus add a new one, how claims remain source- and Artifact-backed, how eval or benchmark evidence stays out of normal project memory, how SkillSearchProcessor and scoped workspace search expose the right guidance on demand, which changes require writing-for-agents and tune-model-evals, and what evidence prevents target-specific hints or benchmark answers from leaking into reusable instructions.

[justsml]

Implemented and pushed the product-skill promotion gate in commit d92a614. It models deepen-existing, new-skill, and retirement paths; requires source-cited reusable claims, a declared independent-campaign threshold, campaign Artifact evidence, contamination and approval-boundary review, evidence-backed validation, and scoped Workspace discovery metadata. Raw campaign/eval/benchmark material cannot be candidate-visible; eval and benchmark inputs are validation-only. The gate emits a reviewed registry plan but performs no filesystem write or runtime activation. Focused integration checks: 16 passing tests including the existing product registry and native Workspace discovery. Full TypeScript check passes. Remaining HITL choices include taxonomy ownership and the promotion threshold, so the ticket stays open.